Skip to main content

Build an Extraction Chain

In this tutorial, we will build a chain to extract structured information from unstructured text.

info

This tutorial will only work with models that support function/tool calling

Concepts

Concepts we will cover are: - Using language models - Using function/tool calling - Debugging and tracing your application using LangSmith

Setup

Installation

To install LangChain run:

yarn add langchain

For more details, see our Installation guide.

LangSmith

Many of the applications you build with LangChain will contain multiple steps with multiple invocations of LLM calls. As these applications get more and more complex, it becomes crucial to be able to inspect what exactly is going on inside your chain or agent. The best way to do this is with LangSmith.

After you sign up at the link above, make sure to set your environment variables to start logging traces:

export LANGCHAIN_TRACING_V2="true"
export LANGCHAIN_API_KEY="..."

The Schema

First, we need to describe what information we want to extract from the text.

We’ll use Zod to define an example schema that extracts personal information.

yarn add zod @langchain/core
import { z } from "zod";

const personSchema = z.object({
name: z.string().nullish().describe("The name of the person"),
hair_color: z
.string()
.nullish()
.describe("The color of the person's hair if known"),
height_in_meters: z.string().nullish().describe("Height measured in meters"),
});

There are two best practices when defining schema:

  1. Document the attributes and the schema itself: This information is sent to the LLM and is used to improve the quality of information extraction.
  2. Do not force the LLM to make up information! Above we used .nullish() for the attributes allowing the LLM to output null or undefined if it doesn’t know the answer.
info

For best performance, document the schema well and make sure the model isn’t force to return results if there’s no information to be extracted in the text.

The Extractor

Let’s create an information extractor using the schema we defined above.

import { ChatPromptTemplate } from "@langchain/core/prompts";

// Define a custom prompt to provide instructions and any additional context.
// 1) You can add examples into the prompt template to improve extraction quality
// 2) Introduce additional parameters to take context into account (e.g., include metadata
// about the document from which the text was extracted.)
const prompt = ChatPromptTemplate.fromMessages([
[
"system",
`You are an expert extraction algorithm.
Only extract relevant information from the text.
If you do not know the value of an attribute asked to extract,
return null for the attribute's value.`,
],
// Please see the how-to about improving performance with
// reference examples.
// ["placeholder", "{examples}"],
["human", "{text}"],
]);

We need to use a model that supports function/tool calling.

Please review the documentation for list of some models that can be used with this API.

import { ChatAnthropic } from "@langchain/anthropic";

const llm = new ChatAnthropic({
model: "claude-3-sonnet-20240229",
temperature: 0,
});

const runnable = prompt.pipe(llm.withStructuredOutput(personSchema));

const text = "Alan Smith is 6 feet tall and has blond hair.";
await runnable.invoke({ text });
{ name: "Alan Smith", hair_color: "blond", height_in_meters: "1.83" }
info

Extraction is Generative 🤯

LLMs are generative models, so they can do some pretty cool things like correctly extract the height of the person in meters even though it was provided in feet!

We can see the LangSmith trace here.

Even though we defined our schema with the variable name personSchema, Zod is unable to infer this name and therefore does not pass it along to the model. To help give the LLM more clues as to what your provided schema represents, you can also give the schema you pass to withStructuredOutput() a name:

const runnable = prompt.pipe(
llm.withStructuredOutput(personSchema, { name: "person" })
);

const text = "Alan Smith is 6 feet tall and has blond hair.";

await runnable.invoke({ text });
{ name: "Alan Smith", hair_color: "blond", height_in_meters: "1.83" }

This can improve performance in many cases.

Multiple Entities

In most cases, you should be extracting a list of entities rather than a single entity.

This can be easily achieved using Zod by nesting models inside one another.

import { z } from "zod";

const personSchema = z.object({
name: z.string().nullish().describe("The name of the person"),
hair_color: z
.string()
.nullish()
.describe("The color of the person's hair if known"),
height_in_meters: z.number().nullish().describe("Height measured in meters"),
});

const dataSchema = z.object({
people: z.array(personSchema).describe("Extracted data about people"),
});
info

Extraction might not be perfect here. Please continue to see how to use Reference Examples to improve the quality of extraction, and see the guidelines section!

const runnable = prompt.pipe(llm.withStructuredOutput(dataSchema));
const text =
"My name is Jeff, my hair is black and i am 6 feet tall. Anna has the same color hair as me.";
await runnable.invoke({ text });
{
people: [
{ name: "Jeff", hair_color: "black", height_in_meters: 1.83 },
{ name: "Anna", hair_color: "black", height_in_meters: null }
]
}
tip

When the schema accommodates the extraction of multiple entities, it also allows the model to extract no entities if no relevant information is in the text by providing an empty list.

This is usually a good thing! It allows specifying required attributes on an entity without necessarily forcing the model to detect this entity.

We can see the LangSmith trace here

Next steps

Now that you understand the basics of extraction with LangChain, you’re ready to proceed to the rest of the how-to guides:

  • Add Examples: Learn how to use reference examples to improve performance.
  • Handle Long Text: What should you do if the text does not fit into the context window of the LLM?
  • Use a Parsing Approach: Use a prompt based approach to extract with models that do not support tool/function calling.

Was this page helpful?


You can leave detailed feedback on GitHub.